作者位次 #2/8ACL 2026Poster· Natural Language Processing
Retrievals Can Be Detrimental: Unveiling the Backdoor Vulnerability of Retrieval-Augmented Diffusion Models
Hao Fang · Xiaohang Sui · Hongyao Yu · Kuofeng Gao · Jiawei Kong · Sijin Yu · Bin Chen · Shu-Tao Xia
摘要
Diffusion models (DMs) have recently exhibited impressive generation capability. However, their training generally requires huge computational resources and large-scale datasets. To solve these, recent studies empower DMs with Retrieval-Augmented Generation (RAG), yielding retrieval-augmented dif...